Improve the performance of the semaphore code

[dave]

Currently, when a semaphore blocks, it has to allocate a task descriptor structure each time. A better approach might be to keep a cache of, say, 16 such structures, so that they can be reused quickly in low contention situations.

It is debatable whether there would be any real performance gain, given that TLSF is a really good memory allocator. The speed difference should be benchmarked.